An Efficient Human-Following Method by Fusing Kernelized Correlation Filter and Depth Information for Mobile Robot

Human-following ability enables mobile robots to cooperate with human beings smartly. The kernelized correlation filter (KCF) was typically adopted as a human tracker during the human-following process due to its advantages of fast speed and high precision. However, it is easy to drift away owing to the changes of human pose. In this study, we proposed an efficient human-following method by fusion of KCF and depth information for the mobile robot. The target human was separated from the background regions based on the seeded region growing. KCF tracker outputs and segmentation results were fused by using the adaptive weighted fusion method. The weighting factors were adjusted based on the tracking quality indicators, which reflect the tracking reliability of KCF and the distribution of depth information. Experimental results verify the effectiveness of the proposed method in terms of tracking speed and stability. The center location error of all frames was less than 48 pixels and the bounding box overlap rate of all frames was larger than 0.62.